In this vLLM Office Hours recording, we dive into batch invariant execution in vLLM and the system and research work enabling it. The session opens with a vLLM project update from Michael Goin, covering recent releases and ecosystem progress, including vLLM v0.11.[1,2] and v0.12.0, vLLM-Omni, vLLM Router, and Speculators v0.3.0. Following the update, Wentao Ye (Red Hat) and Bram Wasti (Meta) walk through the core concepts behind batch invariant execution, explaining what it is, why it matters, and how it is implemented in practice. They then dive into reinforcement learning work built on TorchTitan, covering relevant RL theory, key design implications, and results from recent experiments. This session is aimed at developers and researchers interested in vLLM internals, scalable execution strategies, and reinforcement learning systems. Timestamps 00:00 vLLM project updates and recent release update 21:34 Intro to Batch invariance in vLLM 36:50 Reinforcement learning and TorchTitan deep dive Session slides: https://docs.google.com/presentation/d/1iaZkoyf2VDQFc3TB2DGld2MpZ-uLsVrXdtYWfT9hEBc Explore and register for upcoming vLLM Office Hours: https://red.ht/office-hours

 847      15